Embedded Helicopter Heading Control using an Adaptive Network-Based Fuzzy Inference System

1. Artificial Intelligence Embedded Helicopter Heading Control using an Adaptive Network-Based Fuzzy Inference System Arbab Abdul Waheed MSc in Smart SystemsStudent # 0804515 Nov 17, 2008

2. Introduction Embedded Helicopter Heading Control using an Adaptive Network-Based Fuzzy Inference System

3. Introduction • This paper is about autonomous swarm of small indoor helicopters which is currently being developed at the Centre for Computational Intelligence at the De Montfort University, using high-efficiency and compact embedded control and monitoring • The advantage of the fuzzy approach is that a crisp mathematical model of the system need never be identified • This paper uses TSK (Takagi-Sugeno-Kang) fuzzy interference system implemented on a low-specification embedded hardware board • The FIS is learned from data taken from an existing system using an ANFIS (Adaptive Network-Based Fuzzy Inference System) • ANFIS provides fast and effective means to develop TSK based fuzzy systems close to the original model without the need to formally identify the controllers model

4. Background Embedded Helicopter Heading Control using an Adaptive Network-Based Fuzzy Inference System

5. BACKGROUND • Helicopter model - Twister Bell 47 • The helicopter used in this work is a Twister Bell 47 with twin counter rotating rotors with 340 mm span driven by two high performance DC motors • It also has two servos to control rotor blade angles and a receiver to control the motors and servos over radio control (both not used in this work) • The weight of the helicopter without microcontroller, driver, sensor and batteries is approximately 200 grams • The helicopter has been fixed to a plate on top of a ball bearing, thereby enabling the helicopter to turn freely while not being able to lift off nor tilting to any side • This helicopter doesn’t have a tail rotor of any kind • It comprises of two counter rotating blades is used to give both lift and direction • When both rotors are moving at the same speed a constant heading is maintained • If either rotor‘s speed is reduced the heading will change and lift will be reduced

6. BACKGROUND Helicopter model - Twister Bell 47 • Usually helicopters have 3 rotational degrees of freedom (DOF) called pitch, roll and yaw as shown in figure • They also have 3 translational DOF called up/down, left/right and forwards/backwards

7. BACKGROUND • Microcontroller - OOPIC-R • The heading controller introduced in this work is implemented on a OOPIC-R version B.2.2+ microcontroller • This embedded system has a low power consumption while providing many inputs and outputs great for driving models and robots • The microcontroller has a rather low performance and only 168 Byte of program accessible RAM, it can be programmed in C++, Java and Visual Basic like languages • The microcontroller also provides regulated power supply to the devices connected to it • The microcontroller board has all the technology required except for the external sensor as well as the driver attached to it

8. BACKGROUND • Sensor - Honeywell HMR3000 • The sensor used to read the current heading of the helicopter is the Honeywell HMR3000 Digital Compass Module • This device is highly accurate up to 0.5° • It has a fast response time of up to 20 Hz and a low power consumption of 35 mA • It is communicating over the RS232 using a NMEA (National Marine Electronics Association) protocol • The device needs to be initialised after startup with certain NMEA strings sent over the serial port in order to configure and start the device

9. BACKGROUND • Adaptive Network-Based Fuzzy Inference Systems (ANFIS) • ANFIS is a neuro-fuzzy approach which brings together the machine learning abilities of artificial neural networks with the tractability and robustness of fuzzy logic • ANFIS uses a two pass hybrid learning algorithm to tune these parameters, so that the fuzzy inference system fits the data • For the forward pass, the antecedent parameters are fixed and the consequent parameters are estimated using the least squares estimator • For the backwards pass of the learning algorithm the consequent parameters are now fixed and the antecedent parameters are tuned

10. System Setup Embedded Helicopter Heading Control using an Adaptive Network-Based Fuzzy Inference System

11. System setup • The battery powering the microcontroller and motors is a 7.4 V Li-ion Polymer with 800mAh • The microcontroller board is supplying all electrical devices except for the motors with regulated 5 V • The HMR3000 sensor is connected to the microcontroller over the serial port using a special cable that also provides the sensor with power • The driver is connected to the OOPICs digital motor I/Os as well as to its 5 V power supply • An additional safety switch is connected to the drivers enable input to quickly disable the driver, if needed • An additional and optional TTL logic to serial converter chip is attached to the digital I/Os of the OOPIC to provide a debug interface as the primary serial port is already used by the sensor

12. Implementation Embedded Helicopter Heading Control using an Adaptive Network-Based Fuzzy Inference System

13. implementation • Reading the Sensor • The HMR3000 uses the NMEA protocol to send and receive information • The sensor is configured to transmit new readings using ASCII messages in degrees with 13.75 Hz • A final command initiates the sensor which starts sending information over the serial port, always overwriting old information on the port of the microcontroller • As the information is sent single char wise, one does not know exactly at what point the information is read and analyzed • Therefore synchronization is required and achieved by reading all information until no more information is sent by the sensor • The sensor reading is taken with the following procedure: • Wait for new reading • Read byte by byte into a character array • Generate number out of read information

14. implementation • Proportional Controller • With the heading read from the sensor, the application has now the first and most important of the two inputs needed by the controller • In order to make the heading meaningful to the controller, the difference between the previously fixed and the current heading is calculated by the following equation: • error = current_heading− fixed_heading • The proportional controller reacts based on the following two rules: • The bigger the error, the bigger the turning speed difference • The bigger the total power, the smaller the turning speed difference • Although this is not exactly implemented in a proportional way, the controller is closely related to a proportional controllers and therefore called “proportional controller” in this work

15. implementation • Fuzzy Inference System • As mentioned previously, a TSK fuzzy inference system has been chosen for this work • Such a fuzzy system is usually much less computational expensive than a Mamdani type FIS • It also needs less variable space in terms of RAM which is also extremely small on the microcontroller used in this work • The rules bases for the controller contained six rules • R1 : if μlow(pl) and μlow(e) then w1 • R2 : if μlow(pl) and μmedium(e) then w2 • R3 : if μlow(pl) and μhigh(e) then w3 • R4 : if μhigh(pl) and μlow(e) then w4 • R5 : if μhigh(pl) and μmedium(e) then w5 • R6 : if μhigh(pl) and μhigh(e) then w6

16. implementation • Training parameters with ANFIS • ANFIS uses a hybrid learning approach, the backpropagation gradient descent method to fine-tune the membership functions and the least-squares estimator to identify the consequent parameters

17. implementation • Fixed Point Arithmetic • The microcontroller is not capable of computing floating point numbers, usually needed for the implementation of a FIS • One solution to this problem could be to recreate the FIS to use higher numbers for inputs and output • Rather than changing the design, another method can be used to compute the numbers needed: Fixed Point Arithmetic • Fixed Point Arithmetic has the performance of integer calculations while providing adequate accuracy • As the name of the variable type already states, the point within the variable is fixed • Fixed point variables are created by shifting the value by the fixed point offset to the left

18. implementation • The following example will shows how this is done: • fix1 = 3.1415 << fixed_point_offset; • Using a 2 byte or 16 bit variable and a fixed point offset of 8 bit, the shift operation would be the same as a multiplication by 28−1 = 255 • Thus, a number like 3.1415 would become 801.083 and be stored in the fixed point variable as 801 • Every use and calculation with a fixed point variable needs to be thought through carefully always having in mind the range of possible values in the variables • If the system is designed carefully, this method provides adequate accuracy, is very fast and working well • This method provides the means needed to very efficiently compute membership grades, rules and outputs on the embedded system

19. Results Embedded Helicopter Heading Control using an Adaptive Network-Based Fuzzy Inference System

20. results • The small changes from the previously fixed heading were corrected instantly by slowly rotating the helicopter back to the fixed heading • Medium changes in heading were corrected with more yaw rotation, depending on the total power on the rotors • Large errors between the helicopters and the fixed heading were successfully corrected by the controller although some jiggering appeared • Figure 9 depicts the response graph of the PID controller a moderate change in desired heading of 90o • Figure 10 depicts the response graph of the fuzzy controller a moderate change in desired heading of 90o • Large errors between the helicopters and the fixed heading were successfully corrected by the controller although some jiggering appeared

21. results • In order to measure both controllers cycle times a short “.” message is sent over the debug serial interface after every control cycle • With this method, the control cycle times were measured • The results show that the TSK type fuzzy system used here is only 25% slower than the proportional one • The bottleneck for these results is the OOPIC microcontrollers low performance

22. Conclusion Embedded Helicopter Heading Control using an Adaptive Network-Based Fuzzy Inference System

23. conclusion • It has been shown that the fuzzy logic based controller is performing very well • The speed difference between the proportional and the fuzzy logic based controllers is only about 25% • The error in heading is successfully corrected with only minor problems • The controller has problems with jiggering and in a certain situation when opposite to the previously fixed heading • Unfortunately, the microcontroller used in this work is rather slow and does not have much memory available • Fuzzy logic based controllers are more robust for unstable systems with high variability and uncertainty such as a helicopter • The TSK type fuzzy logic system has been implemented with parameters identified using ANFIS • This method made the development of the fuzzy system easier as no formal controller model had to be identified

24. conclusion • The fuzzy logic based controller is more robust to disturbances and unforeseen inputs • Additionally, it has been shown that using fixed point arithmetic one can implement almost any kind of fuzzy system on an embedded device without the need to modify or tweak the fuzzy system beforehand • In combination with an Adaptive Network-Based Fuzzy Inference System, the TSK type fuzzy system can be implemented based on an existing model without the need to formally identify a control model • Still, an analysis of the system and methods employed showed that there is scope for improvements and future extensions Thank you!